10 - Data Augmentation and Synthetic Data Generation: Growing Datasets

Expand limited data: augmentation for images (rotation, flip, zoom), text (back-translation, paraphrase), time series. SMOTE for imbalanced classification. Generative models (GANs, VAE) for synthetic data. Trade-off: diversity vs realism. When it helps, when it overfits.

What you'll learn

• Geometric transformations: rotation, scaling, flip matrices - the basis of image augmentation
• Mixup: \tilde{x} = \lambda x_i + (1-\lambda) x_j - regularization through interpolation
• SMOTE: interpolates between minority class neighbors to balance
• GAN: min-max game to generate realistic synthetic data
• VAE: reconstruction + KL divergence for a sampleable latent space

This article is part of the Math for AI series on federicocalo.dev.

Read the full article

The complete article (15 min read) with code examples, diagrams, and practical exercises is available here:

➡️ 10 - Data Augmentation and Synthetic Data Generation: Growing Datasets

https://federicocalo.dev/en/blog/data-augmentation-synthetic-data-generation-datasets

By Federico Calò — Software Developer & Technical Writer